By backward I meant the images done by the Deep Dreams program. They retrofit (hence the back) 'dog' signals fractally. This is a pretty amazing primitive to me. Trying to recognize 'concepts' even through that much distortion.
I'd bet a thousands dollars that the underlying principles Deep Dreams uses to infer fractal and continuous images are the same we have in our brains.
That's 100% wrong. It does create abstract representations and it can imagine back. For example, there is this neural net that can do "woman + glasses = woman wearing glasses" on images, without any tags. It can combine concepts and generate new images. There is also another one that can "dream" images of bedrooms. It can even interpolate 10 intermediary bedrooms between any two images, and they look very good.
You absolutely could with a generative adversarial network and the right training dataset. They can interpolate between images in a semantic way, and add and subtract components of the semantic representation. While it's not narwhals and horses, here's an example of using images of men with and without glasses along with a woman without glasses to generate a woman with glasses: https://raw.githubusercontent.com/Newmu/dcgan_code/master/im...
More results here: https://github.com/Newmu/dcgan_code And note that this field is advancing so rapidly that these results have already been improved on despite being only six months old.